DL_SCAN_train test

USER GUIDE

🚀 About this project

This is the project for the Central Weather Bureau (Taiwan) which predicts the initiation of convective cell using Deep learning Model.

Follow the steps to run the model:

💡

• Make sure you prepare the SCAN&Radar data first
• Simply run the pl_run.py to start training
• Model checkpoint will be saved in /root/user/SCAN_checkpoints/
• You can also track the training process by logger which is saved in /root/user/logs with tensorboard
• Utilize the checkpoint to testing the result in another repository /DLSCAN_eval
• Contact r07229013@ntu.edu.tw if you have any problems

🚀

Input data

This is our input data, we use radar CV and convection cell location 1 hour before

Model structure

This is our DLSCAN model schematic diagram. We utilize CNN layer combined with conv-GRU to form a U-NET structure. Then post-process the result with some tricky method (Interesing details will be in our ongoing paper).

Post-processing

This shows one of our post-processing down-scaling method with Maxpool

Final prediction

Here we show the final prediction with and without mask

Done on 15, Spetember, 2022 [name=peterpan]